BlogAI at work

If AI Attends Your Meeting, Give It a Proxy Brief

AI can now sit in on a meeting when you cannot. A proxy brief keeps that convenience from turning your absence into invented authority, silent recording, or a misleading promise.

Jeremy GarciniAug 21, 20266 min read

At 9:02, the project call begins with six people and a seventh tile that belongs to no person in the room. Its owner is double-booked. The bot will take notes for them.

That sounds simple until someone asks, “Would Jeremy be comfortable moving the launch?” The tile cannot read the relationship behind the question. It may not know which dates are flexible, which customer promise is sensitive, or whether silence should mean “no objection.” Yet the group still feels the social pull to treat the absent person as somehow present.

AI meeting attendance is moving from novelty to ordinary workflow. Zoom’s documentation says its AI Companion can join Google Meet and Microsoft Teams as a participant, transcribe the conversation, prepare a summary, and support post-meeting questions. Zoom also provides participant notices, including a labeled tile and chat disclosure. Meanwhile, researchers are testing a more ambitious idea: an AI delegate that can decide when and how to contribute on someone’s behalf.

The useful question is no longer just, “Can a bot attend?” It is, “What exactly is it allowed to represent?”

A note-taker is not yet a delegate

Today’s products vary. Some assistants are essentially visible recorders. Others can retrieve context, suggest actions, or automate follow-through. Research prototypes go further and explore agents that speak for an unavailable person.

Those roles should not blur together. A bot that captures a discussion has observational access. It does not automatically have the owner’s judgment, authority, or consent to make a commitment.

That distinction matters because the technology is still imperfect. In Microsoft Research’s Meeting Delegate benchmark, a prototype was evaluated on real meeting transcripts. About 60 percent of its responses addressed at least one key point from the reference answer, and the researchers still found irrelevant or repetitive contributions and sensitivity to transcription errors. That is promising research, not a basis for pretending an absent colleague has spoken.

A small meeting group working beside an unoccupied chair and a discreet AI meeting device

Before sending AI into a meeting, give it a proxy brief. This is a short agreement shared with the meeting host and, when appropriate, the participants. It defines what the system is there to do and what everyone must leave for the human.

Choose one of three mandates

The fastest way to create confusion is to say that an assistant is “attending for me.” That phrase hides several very different jobs.

Observer: The AI may capture the conversation and return a recap. It cannot ask questions, offer opinions, accept tasks, or signal agreement. This is the right default for most current meeting bots.

Question carrier: The AI may surface a small set of questions prepared by its owner, or flag when a named topic appears. Any improvised contribution should be clearly identified as generated, not presented as the owner’s view.

Bounded delegate: The AI may communicate a documented position within explicit limits, such as “Tuesday or Thursday both work” or “the budget can move by up to five percent.” Anything outside the boundary returns to the human.

Teams should treat these as separate permission levels, not stages a bot can climb on its own. If the invite says observer, a clever inference does not upgrade it to delegate.

Write the brief in six lines

A useful proxy brief should fit beside the agenda. Long instructions will be skipped, and vague instructions will be overinterpreted.

  1. Purpose: Why is the AI present? “Capture decisions and unresolved questions for the product lead.”
  2. Identity: Whom does it support, and how will its presence be disclosed?
  3. Mandate: Observer, question carrier, or bounded delegate.
  4. Priorities: Which two or three topics matter enough to highlight?
  5. No-go zone: What must not be discussed, inferred, or decided through the proxy?
  6. Return path: Which outcomes require human review, by whom, and by when?

For example:

Maya’s AI assistant is joining as an observer because she is with a customer. Please use it only to capture the decision on launch sequencing, risks raised, and questions for Maya. It cannot approve dates, accept work, or represent Maya’s opinion. Any change to customer commitments goes to Maya for confirmation by 3 p.m.

That paragraph does more than constrain software. It gives the room permission to stop itself from manufacturing agreement.

Protect the people who did show up

Consent cannot be reduced to a bot name appearing in the participant list. People need to understand what is being captured, where the output will go, and whether the assistant can act beyond note-taking.

The platform may supply technical notice. Zoom, for example, documents an optional pre-meeting email and in-meeting indicators for its assistant in third-party calls. The meeting owner still carries the social responsibility. Put the purpose in the invitation, ask before enabling capture in a sensitive conversation, and make removal easy without requiring someone to argue a legal case in front of colleagues.

There are also meetings where a proxy is simply the wrong choice. Performance conversations, active conflict, early relationship-building, layoffs, medical or legal matters, and delicate negotiations depend on tone, discretion, and reciprocal attention. A summary can preserve words while losing the fact that showing up was part of the work.

Microsoft Research’s 2026 paper on mimetic proxy agents identifies trust, representation, keeping the human meaningfully informed, and ethical risk as central challenges. That is a useful warning against treating delegated presence as a scheduling trick.

Require a human return receipt

The proxy’s job is not complete when the transcript arrives. The absent person needs a compact route back into responsibility.

A professional reviewing a concise meeting recap with a colleague after returning to the conversation

A good return receipt contains four things:

  • decisions made without the person;
  • commitments proposed in their name;
  • questions that still need their judgment;
  • relationship signals that a transcript may flatten, such as hesitation, disagreement, or a request for a direct conversation.

The human should then confirm, correct, or reject the relevant items. If the group moved beyond the proxy’s mandate, the record should say so plainly: “Provisional until Maya confirms.” This small step prevents a polished recap from laundering an assumption into organizational memory.

It also preserves reciprocity. Colleagues should not have to keep speaking into someone’s automation without ever hearing back from the person. A short response such as “I reviewed the discussion, I confirm the launch date, and I will call Priya about the unresolved risk” closes the loop.

Use AI to support presence, not counterfeit it

Meeting proxies can be genuinely helpful. They can reduce scheduling pressure, preserve context across time zones, and let people protect focus without returning to a blank page. Their value grows when the boundary is obvious.

Sometimes the better choice is still to attend and use AI beside you. Caspi is designed for that version of the problem: real-time recap, suggested questions, contextual chat, proactive flags, action items, and persistent meeting memory while the human remains responsible for the conversation. Whether AI joins in your place or supports you live, the principle is the same. Give the system context, give it limits, and keep authority attached to a person.

Back to all articles